Date of Award
Summer 2010
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Bioinformatics
First Advisor
Struble, Craig A.
Second Advisor
Beardsley, Scott A.
Third Advisor
Ge, Rong
Abstract
Simulating biological neural networks is an important task for computational neuroscientists attempting to model and analyze brain activity and function. As these networks become larger and more complex, the computational power required grows significantly, often requiring the use of supercomputers or compute clusters. An emerging low-cost, highly accessible alternative to many of these resources is the Graphics Processing Unit (GPU) - specialized massively-parallel graphics hardware that has seen increasing use as a general purpose computational accelerator thanks largely due to NVIDIA's CUDA programming interface. We evaluated the relative benefits and limitations of GPU-based tools for large-scale neural network simulation and analysis, first by developing an agent-inspired spiking neural network simulator then by adapting a neural signal decoding algorithm. Under certain network configurations, the simulator was able to outperform an equivalent MPI-based parallel implementation run on a dedicated compute cluster, while the decoding algorithm implementation consistently outperformed its serial counterpart. Additionally, the GPU-based simulator was able to readily visualize network spiking activity in real-time due to the close integration with standard computer graphics APIs. The GPU was shown to provide significant performance benefits under certain circumstances while lagging behind in others. Given the complex nature of these research tasks, a hybrid strategy that combines GPU- and CPU-based approaches provides greater performance than either separately.